cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:06:44 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cfII_32 -dft-II -include rdft/scalar/r2cfII.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 174 FP additions, 128 FP multiplications,
|
cannam@167
|
32 * (or, 46 additions, 0 multiplications, 128 fused multiply/add),
|
cannam@167
|
33 * 62 stack variables, 15 constants, and 64 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/r2cfII.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void r2cfII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
40 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
41 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
42 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
43 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
44 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
45 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
46 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
47 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
48 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
49 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
50 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
51 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
52 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
53 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
54 {
|
cannam@167
|
55 INT i;
|
cannam@167
|
56 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
|
cannam@167
|
57 E T5, T2B, T1z, T2n, Tc, T2C, T1C, T2o, Tm, T1l, T1J, T27, Tv, T1k, T1G;
|
cannam@167
|
58 E T26, T15, T1r, T1Y, T2e, T1c, T1s, T1V, T2d, TK, T1o, T1R, T2b, TR, T1p;
|
cannam@167
|
59 E T1O, T2a;
|
cannam@167
|
60 {
|
cannam@167
|
61 E T1, T2l, T4, T2m, T2, T3;
|
cannam@167
|
62 T1 = R0[0];
|
cannam@167
|
63 T2l = R0[WS(rs, 8)];
|
cannam@167
|
64 T2 = R0[WS(rs, 4)];
|
cannam@167
|
65 T3 = R0[WS(rs, 12)];
|
cannam@167
|
66 T4 = T2 - T3;
|
cannam@167
|
67 T2m = T2 + T3;
|
cannam@167
|
68 T5 = FNMS(KP707106781, T4, T1);
|
cannam@167
|
69 T2B = FNMS(KP707106781, T2m, T2l);
|
cannam@167
|
70 T1z = FMA(KP707106781, T4, T1);
|
cannam@167
|
71 T2n = FMA(KP707106781, T2m, T2l);
|
cannam@167
|
72 }
|
cannam@167
|
73 {
|
cannam@167
|
74 E T8, T1A, Tb, T1B;
|
cannam@167
|
75 {
|
cannam@167
|
76 E T6, T7, T9, Ta;
|
cannam@167
|
77 T6 = R0[WS(rs, 10)];
|
cannam@167
|
78 T7 = R0[WS(rs, 2)];
|
cannam@167
|
79 T8 = FMA(KP414213562, T7, T6);
|
cannam@167
|
80 T1A = FNMS(KP414213562, T6, T7);
|
cannam@167
|
81 T9 = R0[WS(rs, 6)];
|
cannam@167
|
82 Ta = R0[WS(rs, 14)];
|
cannam@167
|
83 Tb = FMA(KP414213562, Ta, T9);
|
cannam@167
|
84 T1B = FMS(KP414213562, T9, Ta);
|
cannam@167
|
85 }
|
cannam@167
|
86 Tc = T8 - Tb;
|
cannam@167
|
87 T2C = T1B - T1A;
|
cannam@167
|
88 T1C = T1A + T1B;
|
cannam@167
|
89 T2o = T8 + Tb;
|
cannam@167
|
90 }
|
cannam@167
|
91 {
|
cannam@167
|
92 E Te, Tj, Th, Tk, Tf, Tg;
|
cannam@167
|
93 Te = R0[WS(rs, 7)];
|
cannam@167
|
94 Tj = R0[WS(rs, 15)];
|
cannam@167
|
95 Tf = R0[WS(rs, 3)];
|
cannam@167
|
96 Tg = R0[WS(rs, 11)];
|
cannam@167
|
97 Th = Tf + Tg;
|
cannam@167
|
98 Tk = Tg - Tf;
|
cannam@167
|
99 {
|
cannam@167
|
100 E Ti, Tl, T1H, T1I;
|
cannam@167
|
101 Ti = FNMS(KP707106781, Th, Te);
|
cannam@167
|
102 Tl = FNMS(KP707106781, Tk, Tj);
|
cannam@167
|
103 Tm = FNMS(KP668178637, Tl, Ti);
|
cannam@167
|
104 T1l = FMA(KP668178637, Ti, Tl);
|
cannam@167
|
105 T1H = FMA(KP707106781, Th, Te);
|
cannam@167
|
106 T1I = FMA(KP707106781, Tk, Tj);
|
cannam@167
|
107 T1J = FMA(KP198912367, T1I, T1H);
|
cannam@167
|
108 T27 = FNMS(KP198912367, T1H, T1I);
|
cannam@167
|
109 }
|
cannam@167
|
110 }
|
cannam@167
|
111 {
|
cannam@167
|
112 E Tn, Ts, Tq, Tt, To, Tp;
|
cannam@167
|
113 Tn = R0[WS(rs, 9)];
|
cannam@167
|
114 Ts = R0[WS(rs, 1)];
|
cannam@167
|
115 To = R0[WS(rs, 5)];
|
cannam@167
|
116 Tp = R0[WS(rs, 13)];
|
cannam@167
|
117 Tq = To + Tp;
|
cannam@167
|
118 Tt = To - Tp;
|
cannam@167
|
119 {
|
cannam@167
|
120 E Tr, Tu, T1E, T1F;
|
cannam@167
|
121 Tr = FNMS(KP707106781, Tq, Tn);
|
cannam@167
|
122 Tu = FNMS(KP707106781, Tt, Ts);
|
cannam@167
|
123 Tv = FNMS(KP668178637, Tu, Tr);
|
cannam@167
|
124 T1k = FMA(KP668178637, Tr, Tu);
|
cannam@167
|
125 T1E = FMA(KP707106781, Tq, Tn);
|
cannam@167
|
126 T1F = FMA(KP707106781, Tt, Ts);
|
cannam@167
|
127 T1G = FMA(KP198912367, T1F, T1E);
|
cannam@167
|
128 T26 = FNMS(KP198912367, T1E, T1F);
|
cannam@167
|
129 }
|
cannam@167
|
130 }
|
cannam@167
|
131 {
|
cannam@167
|
132 E TT, T16, TW, T17, T10, T1a, T13, T19, TU, TV;
|
cannam@167
|
133 TT = R1[WS(rs, 15)];
|
cannam@167
|
134 T16 = R1[WS(rs, 7)];
|
cannam@167
|
135 TU = R1[WS(rs, 3)];
|
cannam@167
|
136 TV = R1[WS(rs, 11)];
|
cannam@167
|
137 TW = TU - TV;
|
cannam@167
|
138 T17 = TU + TV;
|
cannam@167
|
139 {
|
cannam@167
|
140 E TY, TZ, T11, T12;
|
cannam@167
|
141 TY = R1[WS(rs, 9)];
|
cannam@167
|
142 TZ = R1[WS(rs, 1)];
|
cannam@167
|
143 T10 = FMA(KP414213562, TZ, TY);
|
cannam@167
|
144 T1a = FNMS(KP414213562, TY, TZ);
|
cannam@167
|
145 T11 = R1[WS(rs, 5)];
|
cannam@167
|
146 T12 = R1[WS(rs, 13)];
|
cannam@167
|
147 T13 = FMA(KP414213562, T12, T11);
|
cannam@167
|
148 T19 = FMS(KP414213562, T11, T12);
|
cannam@167
|
149 }
|
cannam@167
|
150 {
|
cannam@167
|
151 E TX, T14, T1W, T1X;
|
cannam@167
|
152 TX = FMA(KP707106781, TW, TT);
|
cannam@167
|
153 T14 = T10 - T13;
|
cannam@167
|
154 T15 = FMA(KP923879532, T14, TX);
|
cannam@167
|
155 T1r = FNMS(KP923879532, T14, TX);
|
cannam@167
|
156 T1W = FMA(KP707106781, T17, T16);
|
cannam@167
|
157 T1X = T10 + T13;
|
cannam@167
|
158 T1Y = FNMS(KP923879532, T1X, T1W);
|
cannam@167
|
159 T2e = FMA(KP923879532, T1X, T1W);
|
cannam@167
|
160 }
|
cannam@167
|
161 {
|
cannam@167
|
162 E T18, T1b, T1T, T1U;
|
cannam@167
|
163 T18 = FNMS(KP707106781, T17, T16);
|
cannam@167
|
164 T1b = T19 - T1a;
|
cannam@167
|
165 T1c = FNMS(KP923879532, T1b, T18);
|
cannam@167
|
166 T1s = FMA(KP923879532, T1b, T18);
|
cannam@167
|
167 T1T = FMS(KP707106781, TW, TT);
|
cannam@167
|
168 T1U = T1a + T19;
|
cannam@167
|
169 T1V = FNMS(KP923879532, T1U, T1T);
|
cannam@167
|
170 T2d = FMA(KP923879532, T1U, T1T);
|
cannam@167
|
171 }
|
cannam@167
|
172 }
|
cannam@167
|
173 {
|
cannam@167
|
174 E Ty, TL, TB, TM, TF, TP, TI, TO, Tz, TA;
|
cannam@167
|
175 Ty = R1[0];
|
cannam@167
|
176 TL = R1[WS(rs, 8)];
|
cannam@167
|
177 Tz = R1[WS(rs, 4)];
|
cannam@167
|
178 TA = R1[WS(rs, 12)];
|
cannam@167
|
179 TB = Tz - TA;
|
cannam@167
|
180 TM = Tz + TA;
|
cannam@167
|
181 {
|
cannam@167
|
182 E TD, TE, TG, TH;
|
cannam@167
|
183 TD = R1[WS(rs, 10)];
|
cannam@167
|
184 TE = R1[WS(rs, 2)];
|
cannam@167
|
185 TF = FMA(KP414213562, TE, TD);
|
cannam@167
|
186 TP = FNMS(KP414213562, TD, TE);
|
cannam@167
|
187 TG = R1[WS(rs, 6)];
|
cannam@167
|
188 TH = R1[WS(rs, 14)];
|
cannam@167
|
189 TI = FMA(KP414213562, TH, TG);
|
cannam@167
|
190 TO = FMS(KP414213562, TG, TH);
|
cannam@167
|
191 }
|
cannam@167
|
192 {
|
cannam@167
|
193 E TC, TJ, T1P, T1Q;
|
cannam@167
|
194 TC = FNMS(KP707106781, TB, Ty);
|
cannam@167
|
195 TJ = TF - TI;
|
cannam@167
|
196 TK = FNMS(KP923879532, TJ, TC);
|
cannam@167
|
197 T1o = FMA(KP923879532, TJ, TC);
|
cannam@167
|
198 T1P = FMA(KP707106781, TM, TL);
|
cannam@167
|
199 T1Q = TF + TI;
|
cannam@167
|
200 T1R = FNMS(KP923879532, T1Q, T1P);
|
cannam@167
|
201 T2b = FMA(KP923879532, T1Q, T1P);
|
cannam@167
|
202 }
|
cannam@167
|
203 {
|
cannam@167
|
204 E TN, TQ, T1M, T1N;
|
cannam@167
|
205 TN = FNMS(KP707106781, TM, TL);
|
cannam@167
|
206 TQ = TO - TP;
|
cannam@167
|
207 TR = FNMS(KP923879532, TQ, TN);
|
cannam@167
|
208 T1p = FMA(KP923879532, TQ, TN);
|
cannam@167
|
209 T1M = FMA(KP707106781, TB, Ty);
|
cannam@167
|
210 T1N = TP + TO;
|
cannam@167
|
211 T1O = FNMS(KP923879532, T1N, T1M);
|
cannam@167
|
212 T2a = FMA(KP923879532, T1N, T1M);
|
cannam@167
|
213 }
|
cannam@167
|
214 }
|
cannam@167
|
215 {
|
cannam@167
|
216 E Tx, T1f, T2L, T2N, T1e, T2O, T1i, T2M;
|
cannam@167
|
217 {
|
cannam@167
|
218 E Td, Tw, T2J, T2K;
|
cannam@167
|
219 Td = FNMS(KP923879532, Tc, T5);
|
cannam@167
|
220 Tw = Tm - Tv;
|
cannam@167
|
221 Tx = FMA(KP831469612, Tw, Td);
|
cannam@167
|
222 T1f = FNMS(KP831469612, Tw, Td);
|
cannam@167
|
223 T2J = FNMS(KP923879532, T2C, T2B);
|
cannam@167
|
224 T2K = T1k + T1l;
|
cannam@167
|
225 T2L = FMA(KP831469612, T2K, T2J);
|
cannam@167
|
226 T2N = FNMS(KP831469612, T2K, T2J);
|
cannam@167
|
227 }
|
cannam@167
|
228 {
|
cannam@167
|
229 E TS, T1d, T1g, T1h;
|
cannam@167
|
230 TS = FNMS(KP534511135, TR, TK);
|
cannam@167
|
231 T1d = FNMS(KP534511135, T1c, T15);
|
cannam@167
|
232 T1e = TS - T1d;
|
cannam@167
|
233 T2O = TS + T1d;
|
cannam@167
|
234 T1g = FMA(KP534511135, TK, TR);
|
cannam@167
|
235 T1h = FMA(KP534511135, T15, T1c);
|
cannam@167
|
236 T1i = T1g - T1h;
|
cannam@167
|
237 T2M = T1g + T1h;
|
cannam@167
|
238 }
|
cannam@167
|
239 Cr[WS(csr, 13)] = FNMS(KP881921264, T1e, Tx);
|
cannam@167
|
240 Ci[WS(csi, 13)] = FNMS(KP881921264, T2M, T2L);
|
cannam@167
|
241 Cr[WS(csr, 2)] = FMA(KP881921264, T1e, Tx);
|
cannam@167
|
242 Ci[WS(csi, 2)] = -(FMA(KP881921264, T2M, T2L));
|
cannam@167
|
243 Cr[WS(csr, 10)] = FNMS(KP881921264, T1i, T1f);
|
cannam@167
|
244 Ci[WS(csi, 10)] = -(FMA(KP881921264, T2O, T2N));
|
cannam@167
|
245 Cr[WS(csr, 5)] = FMA(KP881921264, T1i, T1f);
|
cannam@167
|
246 Ci[WS(csi, 5)] = FNMS(KP881921264, T2O, T2N);
|
cannam@167
|
247 }
|
cannam@167
|
248 {
|
cannam@167
|
249 E T29, T2h, T2r, T2t, T2g, T2u, T2k, T2s;
|
cannam@167
|
250 {
|
cannam@167
|
251 E T25, T28, T2p, T2q;
|
cannam@167
|
252 T25 = FMA(KP923879532, T1C, T1z);
|
cannam@167
|
253 T28 = T26 - T27;
|
cannam@167
|
254 T29 = FMA(KP980785280, T28, T25);
|
cannam@167
|
255 T2h = FNMS(KP980785280, T28, T25);
|
cannam@167
|
256 T2p = FMA(KP923879532, T2o, T2n);
|
cannam@167
|
257 T2q = T1G + T1J;
|
cannam@167
|
258 T2r = FMA(KP980785280, T2q, T2p);
|
cannam@167
|
259 T2t = FNMS(KP980785280, T2q, T2p);
|
cannam@167
|
260 }
|
cannam@167
|
261 {
|
cannam@167
|
262 E T2c, T2f, T2i, T2j;
|
cannam@167
|
263 T2c = FNMS(KP098491403, T2b, T2a);
|
cannam@167
|
264 T2f = FMA(KP098491403, T2e, T2d);
|
cannam@167
|
265 T2g = T2c + T2f;
|
cannam@167
|
266 T2u = T2f - T2c;
|
cannam@167
|
267 T2i = FMA(KP098491403, T2a, T2b);
|
cannam@167
|
268 T2j = FNMS(KP098491403, T2d, T2e);
|
cannam@167
|
269 T2k = T2i - T2j;
|
cannam@167
|
270 T2s = T2i + T2j;
|
cannam@167
|
271 }
|
cannam@167
|
272 Cr[WS(csr, 15)] = FNMS(KP995184726, T2g, T29);
|
cannam@167
|
273 Ci[WS(csi, 15)] = FNMS(KP995184726, T2s, T2r);
|
cannam@167
|
274 Cr[0] = FMA(KP995184726, T2g, T29);
|
cannam@167
|
275 Ci[0] = -(FMA(KP995184726, T2s, T2r));
|
cannam@167
|
276 Cr[WS(csr, 8)] = FNMS(KP995184726, T2k, T2h);
|
cannam@167
|
277 Ci[WS(csi, 8)] = FMS(KP995184726, T2u, T2t);
|
cannam@167
|
278 Cr[WS(csr, 7)] = FMA(KP995184726, T2k, T2h);
|
cannam@167
|
279 Ci[WS(csi, 7)] = FMA(KP995184726, T2u, T2t);
|
cannam@167
|
280 }
|
cannam@167
|
281 {
|
cannam@167
|
282 E T1n, T1v, T2F, T2H, T1u, T2I, T1y, T2G;
|
cannam@167
|
283 {
|
cannam@167
|
284 E T1j, T1m, T2D, T2E;
|
cannam@167
|
285 T1j = FMA(KP923879532, Tc, T5);
|
cannam@167
|
286 T1m = T1k - T1l;
|
cannam@167
|
287 T1n = FMA(KP831469612, T1m, T1j);
|
cannam@167
|
288 T1v = FNMS(KP831469612, T1m, T1j);
|
cannam@167
|
289 T2D = FMA(KP923879532, T2C, T2B);
|
cannam@167
|
290 T2E = Tv + Tm;
|
cannam@167
|
291 T2F = FMA(KP831469612, T2E, T2D);
|
cannam@167
|
292 T2H = FNMS(KP831469612, T2E, T2D);
|
cannam@167
|
293 }
|
cannam@167
|
294 {
|
cannam@167
|
295 E T1q, T1t, T1w, T1x;
|
cannam@167
|
296 T1q = FMA(KP303346683, T1p, T1o);
|
cannam@167
|
297 T1t = FMA(KP303346683, T1s, T1r);
|
cannam@167
|
298 T1u = T1q - T1t;
|
cannam@167
|
299 T2I = T1q + T1t;
|
cannam@167
|
300 T1w = FNMS(KP303346683, T1r, T1s);
|
cannam@167
|
301 T1x = FNMS(KP303346683, T1o, T1p);
|
cannam@167
|
302 T1y = T1w - T1x;
|
cannam@167
|
303 T2G = T1x + T1w;
|
cannam@167
|
304 }
|
cannam@167
|
305 Cr[WS(csr, 14)] = FNMS(KP956940335, T1u, T1n);
|
cannam@167
|
306 Ci[WS(csi, 14)] = FMS(KP956940335, T2G, T2F);
|
cannam@167
|
307 Cr[WS(csr, 1)] = FMA(KP956940335, T1u, T1n);
|
cannam@167
|
308 Ci[WS(csi, 1)] = FMA(KP956940335, T2G, T2F);
|
cannam@167
|
309 Cr[WS(csr, 9)] = FNMS(KP956940335, T1y, T1v);
|
cannam@167
|
310 Ci[WS(csi, 9)] = FNMS(KP956940335, T2I, T2H);
|
cannam@167
|
311 Cr[WS(csr, 6)] = FMA(KP956940335, T1y, T1v);
|
cannam@167
|
312 Ci[WS(csi, 6)] = -(FMA(KP956940335, T2I, T2H));
|
cannam@167
|
313 }
|
cannam@167
|
314 {
|
cannam@167
|
315 E T1L, T21, T2x, T2z, T20, T2A, T24, T2y;
|
cannam@167
|
316 {
|
cannam@167
|
317 E T1D, T1K, T2v, T2w;
|
cannam@167
|
318 T1D = FNMS(KP923879532, T1C, T1z);
|
cannam@167
|
319 T1K = T1G - T1J;
|
cannam@167
|
320 T1L = FMA(KP980785280, T1K, T1D);
|
cannam@167
|
321 T21 = FNMS(KP980785280, T1K, T1D);
|
cannam@167
|
322 T2v = FNMS(KP923879532, T2o, T2n);
|
cannam@167
|
323 T2w = T26 + T27;
|
cannam@167
|
324 T2x = FNMS(KP980785280, T2w, T2v);
|
cannam@167
|
325 T2z = FMA(KP980785280, T2w, T2v);
|
cannam@167
|
326 }
|
cannam@167
|
327 {
|
cannam@167
|
328 E T1S, T1Z, T22, T23;
|
cannam@167
|
329 T1S = FMA(KP820678790, T1R, T1O);
|
cannam@167
|
330 T1Z = FNMS(KP820678790, T1Y, T1V);
|
cannam@167
|
331 T20 = T1S + T1Z;
|
cannam@167
|
332 T2A = T1Z - T1S;
|
cannam@167
|
333 T22 = FMA(KP820678790, T1V, T1Y);
|
cannam@167
|
334 T23 = FNMS(KP820678790, T1O, T1R);
|
cannam@167
|
335 T24 = T22 - T23;
|
cannam@167
|
336 T2y = T23 + T22;
|
cannam@167
|
337 }
|
cannam@167
|
338 Cr[WS(csr, 12)] = FNMS(KP773010453, T20, T1L);
|
cannam@167
|
339 Ci[WS(csi, 12)] = FMS(KP773010453, T2y, T2x);
|
cannam@167
|
340 Cr[WS(csr, 3)] = FMA(KP773010453, T20, T1L);
|
cannam@167
|
341 Ci[WS(csi, 3)] = FMA(KP773010453, T2y, T2x);
|
cannam@167
|
342 Cr[WS(csr, 11)] = FNMS(KP773010453, T24, T21);
|
cannam@167
|
343 Ci[WS(csi, 11)] = FMA(KP773010453, T2A, T2z);
|
cannam@167
|
344 Cr[WS(csr, 4)] = FMA(KP773010453, T24, T21);
|
cannam@167
|
345 Ci[WS(csi, 4)] = FMS(KP773010453, T2A, T2z);
|
cannam@167
|
346 }
|
cannam@167
|
347 }
|
cannam@167
|
348 }
|
cannam@167
|
349 }
|
cannam@167
|
350
|
cannam@167
|
351 static const kr2c_desc desc = { 32, "r2cfII_32", {46, 0, 128, 0}, &GENUS };
|
cannam@167
|
352
|
cannam@167
|
353 void X(codelet_r2cfII_32) (planner *p) {
|
cannam@167
|
354 X(kr2c_register) (p, r2cfII_32, &desc);
|
cannam@167
|
355 }
|
cannam@167
|
356
|
cannam@167
|
357 #else
|
cannam@167
|
358
|
cannam@167
|
359 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cfII_32 -dft-II -include rdft/scalar/r2cfII.h */
|
cannam@167
|
360
|
cannam@167
|
361 /*
|
cannam@167
|
362 * This function contains 174 FP additions, 82 FP multiplications,
|
cannam@167
|
363 * (or, 138 additions, 46 multiplications, 36 fused multiply/add),
|
cannam@167
|
364 * 62 stack variables, 15 constants, and 64 memory accesses
|
cannam@167
|
365 */
|
cannam@167
|
366 #include "rdft/scalar/r2cfII.h"
|
cannam@167
|
367
|
cannam@167
|
368 static void r2cfII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
369 {
|
cannam@167
|
370 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
cannam@167
|
371 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
372 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
cannam@167
|
373 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
374 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
cannam@167
|
375 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
376 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
377 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
cannam@167
|
378 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
379 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
380 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
381 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
382 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
383 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
384 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
385 {
|
cannam@167
|
386 INT i;
|
cannam@167
|
387 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
|
cannam@167
|
388 E T5, T2D, T1z, T2q, Tc, T2C, T1C, T2n, Tm, T1k, T1J, T26, Tv, T1l, T1G;
|
cannam@167
|
389 E T27, T15, T1r, T1Y, T2e, T1c, T1s, T1V, T2d, TK, T1o, T1R, T2b, TR, T1p;
|
cannam@167
|
390 E T1O, T2a;
|
cannam@167
|
391 {
|
cannam@167
|
392 E T1, T2p, T4, T2o, T2, T3;
|
cannam@167
|
393 T1 = R0[0];
|
cannam@167
|
394 T2p = R0[WS(rs, 8)];
|
cannam@167
|
395 T2 = R0[WS(rs, 4)];
|
cannam@167
|
396 T3 = R0[WS(rs, 12)];
|
cannam@167
|
397 T4 = KP707106781 * (T2 - T3);
|
cannam@167
|
398 T2o = KP707106781 * (T2 + T3);
|
cannam@167
|
399 T5 = T1 + T4;
|
cannam@167
|
400 T2D = T2p - T2o;
|
cannam@167
|
401 T1z = T1 - T4;
|
cannam@167
|
402 T2q = T2o + T2p;
|
cannam@167
|
403 }
|
cannam@167
|
404 {
|
cannam@167
|
405 E T8, T1A, Tb, T1B;
|
cannam@167
|
406 {
|
cannam@167
|
407 E T6, T7, T9, Ta;
|
cannam@167
|
408 T6 = R0[WS(rs, 2)];
|
cannam@167
|
409 T7 = R0[WS(rs, 10)];
|
cannam@167
|
410 T8 = FNMS(KP382683432, T7, KP923879532 * T6);
|
cannam@167
|
411 T1A = FMA(KP382683432, T6, KP923879532 * T7);
|
cannam@167
|
412 T9 = R0[WS(rs, 6)];
|
cannam@167
|
413 Ta = R0[WS(rs, 14)];
|
cannam@167
|
414 Tb = FNMS(KP923879532, Ta, KP382683432 * T9);
|
cannam@167
|
415 T1B = FMA(KP923879532, T9, KP382683432 * Ta);
|
cannam@167
|
416 }
|
cannam@167
|
417 Tc = T8 + Tb;
|
cannam@167
|
418 T2C = Tb - T8;
|
cannam@167
|
419 T1C = T1A - T1B;
|
cannam@167
|
420 T2n = T1A + T1B;
|
cannam@167
|
421 }
|
cannam@167
|
422 {
|
cannam@167
|
423 E Te, Tk, Th, Tj, Tf, Tg;
|
cannam@167
|
424 Te = R0[WS(rs, 1)];
|
cannam@167
|
425 Tk = R0[WS(rs, 9)];
|
cannam@167
|
426 Tf = R0[WS(rs, 5)];
|
cannam@167
|
427 Tg = R0[WS(rs, 13)];
|
cannam@167
|
428 Th = KP707106781 * (Tf - Tg);
|
cannam@167
|
429 Tj = KP707106781 * (Tf + Tg);
|
cannam@167
|
430 {
|
cannam@167
|
431 E Ti, Tl, T1H, T1I;
|
cannam@167
|
432 Ti = Te + Th;
|
cannam@167
|
433 Tl = Tj + Tk;
|
cannam@167
|
434 Tm = FNMS(KP195090322, Tl, KP980785280 * Ti);
|
cannam@167
|
435 T1k = FMA(KP195090322, Ti, KP980785280 * Tl);
|
cannam@167
|
436 T1H = Tk - Tj;
|
cannam@167
|
437 T1I = Te - Th;
|
cannam@167
|
438 T1J = FNMS(KP555570233, T1I, KP831469612 * T1H);
|
cannam@167
|
439 T26 = FMA(KP831469612, T1I, KP555570233 * T1H);
|
cannam@167
|
440 }
|
cannam@167
|
441 }
|
cannam@167
|
442 {
|
cannam@167
|
443 E Tq, Tt, Tp, Ts, Tn, To;
|
cannam@167
|
444 Tq = R0[WS(rs, 15)];
|
cannam@167
|
445 Tt = R0[WS(rs, 7)];
|
cannam@167
|
446 Tn = R0[WS(rs, 3)];
|
cannam@167
|
447 To = R0[WS(rs, 11)];
|
cannam@167
|
448 Tp = KP707106781 * (Tn - To);
|
cannam@167
|
449 Ts = KP707106781 * (Tn + To);
|
cannam@167
|
450 {
|
cannam@167
|
451 E Tr, Tu, T1E, T1F;
|
cannam@167
|
452 Tr = Tp - Tq;
|
cannam@167
|
453 Tu = Ts + Tt;
|
cannam@167
|
454 Tv = FMA(KP980785280, Tr, KP195090322 * Tu);
|
cannam@167
|
455 T1l = FNMS(KP980785280, Tu, KP195090322 * Tr);
|
cannam@167
|
456 T1E = Tt - Ts;
|
cannam@167
|
457 T1F = Tp + Tq;
|
cannam@167
|
458 T1G = FNMS(KP555570233, T1F, KP831469612 * T1E);
|
cannam@167
|
459 T27 = FMA(KP831469612, T1F, KP555570233 * T1E);
|
cannam@167
|
460 }
|
cannam@167
|
461 }
|
cannam@167
|
462 {
|
cannam@167
|
463 E TW, T1a, TV, T19, T10, T16, T13, T17, TT, TU;
|
cannam@167
|
464 TW = R1[WS(rs, 15)];
|
cannam@167
|
465 T1a = R1[WS(rs, 7)];
|
cannam@167
|
466 TT = R1[WS(rs, 3)];
|
cannam@167
|
467 TU = R1[WS(rs, 11)];
|
cannam@167
|
468 TV = KP707106781 * (TT - TU);
|
cannam@167
|
469 T19 = KP707106781 * (TT + TU);
|
cannam@167
|
470 {
|
cannam@167
|
471 E TY, TZ, T11, T12;
|
cannam@167
|
472 TY = R1[WS(rs, 1)];
|
cannam@167
|
473 TZ = R1[WS(rs, 9)];
|
cannam@167
|
474 T10 = FNMS(KP382683432, TZ, KP923879532 * TY);
|
cannam@167
|
475 T16 = FMA(KP382683432, TY, KP923879532 * TZ);
|
cannam@167
|
476 T11 = R1[WS(rs, 5)];
|
cannam@167
|
477 T12 = R1[WS(rs, 13)];
|
cannam@167
|
478 T13 = FNMS(KP923879532, T12, KP382683432 * T11);
|
cannam@167
|
479 T17 = FMA(KP923879532, T11, KP382683432 * T12);
|
cannam@167
|
480 }
|
cannam@167
|
481 {
|
cannam@167
|
482 E TX, T14, T1W, T1X;
|
cannam@167
|
483 TX = TV - TW;
|
cannam@167
|
484 T14 = T10 + T13;
|
cannam@167
|
485 T15 = TX + T14;
|
cannam@167
|
486 T1r = TX - T14;
|
cannam@167
|
487 T1W = T13 - T10;
|
cannam@167
|
488 T1X = T1a - T19;
|
cannam@167
|
489 T1Y = T1W - T1X;
|
cannam@167
|
490 T2e = T1W + T1X;
|
cannam@167
|
491 }
|
cannam@167
|
492 {
|
cannam@167
|
493 E T18, T1b, T1T, T1U;
|
cannam@167
|
494 T18 = T16 + T17;
|
cannam@167
|
495 T1b = T19 + T1a;
|
cannam@167
|
496 T1c = T18 + T1b;
|
cannam@167
|
497 T1s = T1b - T18;
|
cannam@167
|
498 T1T = TV + TW;
|
cannam@167
|
499 T1U = T16 - T17;
|
cannam@167
|
500 T1V = T1T + T1U;
|
cannam@167
|
501 T2d = T1U - T1T;
|
cannam@167
|
502 }
|
cannam@167
|
503 }
|
cannam@167
|
504 {
|
cannam@167
|
505 E Ty, TP, TB, TO, TF, TL, TI, TM, Tz, TA;
|
cannam@167
|
506 Ty = R1[0];
|
cannam@167
|
507 TP = R1[WS(rs, 8)];
|
cannam@167
|
508 Tz = R1[WS(rs, 4)];
|
cannam@167
|
509 TA = R1[WS(rs, 12)];
|
cannam@167
|
510 TB = KP707106781 * (Tz - TA);
|
cannam@167
|
511 TO = KP707106781 * (Tz + TA);
|
cannam@167
|
512 {
|
cannam@167
|
513 E TD, TE, TG, TH;
|
cannam@167
|
514 TD = R1[WS(rs, 2)];
|
cannam@167
|
515 TE = R1[WS(rs, 10)];
|
cannam@167
|
516 TF = FNMS(KP382683432, TE, KP923879532 * TD);
|
cannam@167
|
517 TL = FMA(KP382683432, TD, KP923879532 * TE);
|
cannam@167
|
518 TG = R1[WS(rs, 6)];
|
cannam@167
|
519 TH = R1[WS(rs, 14)];
|
cannam@167
|
520 TI = FNMS(KP923879532, TH, KP382683432 * TG);
|
cannam@167
|
521 TM = FMA(KP923879532, TG, KP382683432 * TH);
|
cannam@167
|
522 }
|
cannam@167
|
523 {
|
cannam@167
|
524 E TC, TJ, T1P, T1Q;
|
cannam@167
|
525 TC = Ty + TB;
|
cannam@167
|
526 TJ = TF + TI;
|
cannam@167
|
527 TK = TC + TJ;
|
cannam@167
|
528 T1o = TC - TJ;
|
cannam@167
|
529 T1P = TI - TF;
|
cannam@167
|
530 T1Q = TP - TO;
|
cannam@167
|
531 T1R = T1P - T1Q;
|
cannam@167
|
532 T2b = T1P + T1Q;
|
cannam@167
|
533 }
|
cannam@167
|
534 {
|
cannam@167
|
535 E TN, TQ, T1M, T1N;
|
cannam@167
|
536 TN = TL + TM;
|
cannam@167
|
537 TQ = TO + TP;
|
cannam@167
|
538 TR = TN + TQ;
|
cannam@167
|
539 T1p = TQ - TN;
|
cannam@167
|
540 T1M = Ty - TB;
|
cannam@167
|
541 T1N = TL - TM;
|
cannam@167
|
542 T1O = T1M - T1N;
|
cannam@167
|
543 T2a = T1M + T1N;
|
cannam@167
|
544 }
|
cannam@167
|
545 }
|
cannam@167
|
546 {
|
cannam@167
|
547 E Tx, T1f, T2s, T2u, T1e, T2l, T1i, T2t;
|
cannam@167
|
548 {
|
cannam@167
|
549 E Td, Tw, T2m, T2r;
|
cannam@167
|
550 Td = T5 + Tc;
|
cannam@167
|
551 Tw = Tm + Tv;
|
cannam@167
|
552 Tx = Td - Tw;
|
cannam@167
|
553 T1f = Td + Tw;
|
cannam@167
|
554 T2m = T1l - T1k;
|
cannam@167
|
555 T2r = T2n + T2q;
|
cannam@167
|
556 T2s = T2m - T2r;
|
cannam@167
|
557 T2u = T2m + T2r;
|
cannam@167
|
558 }
|
cannam@167
|
559 {
|
cannam@167
|
560 E TS, T1d, T1g, T1h;
|
cannam@167
|
561 TS = FMA(KP098017140, TK, KP995184726 * TR);
|
cannam@167
|
562 T1d = FNMS(KP995184726, T1c, KP098017140 * T15);
|
cannam@167
|
563 T1e = TS + T1d;
|
cannam@167
|
564 T2l = T1d - TS;
|
cannam@167
|
565 T1g = FNMS(KP098017140, TR, KP995184726 * TK);
|
cannam@167
|
566 T1h = FMA(KP995184726, T15, KP098017140 * T1c);
|
cannam@167
|
567 T1i = T1g + T1h;
|
cannam@167
|
568 T2t = T1h - T1g;
|
cannam@167
|
569 }
|
cannam@167
|
570 Cr[WS(csr, 8)] = Tx - T1e;
|
cannam@167
|
571 Ci[WS(csi, 8)] = T2t - T2u;
|
cannam@167
|
572 Cr[WS(csr, 7)] = Tx + T1e;
|
cannam@167
|
573 Ci[WS(csi, 7)] = T2t + T2u;
|
cannam@167
|
574 Cr[WS(csr, 15)] = T1f - T1i;
|
cannam@167
|
575 Ci[WS(csi, 15)] = T2l - T2s;
|
cannam@167
|
576 Cr[0] = T1f + T1i;
|
cannam@167
|
577 Ci[0] = T2l + T2s;
|
cannam@167
|
578 }
|
cannam@167
|
579 {
|
cannam@167
|
580 E T29, T2h, T2M, T2O, T2g, T2J, T2k, T2N;
|
cannam@167
|
581 {
|
cannam@167
|
582 E T25, T28, T2K, T2L;
|
cannam@167
|
583 T25 = T1z + T1C;
|
cannam@167
|
584 T28 = T26 - T27;
|
cannam@167
|
585 T29 = T25 + T28;
|
cannam@167
|
586 T2h = T25 - T28;
|
cannam@167
|
587 T2K = T1J + T1G;
|
cannam@167
|
588 T2L = T2C + T2D;
|
cannam@167
|
589 T2M = T2K - T2L;
|
cannam@167
|
590 T2O = T2K + T2L;
|
cannam@167
|
591 }
|
cannam@167
|
592 {
|
cannam@167
|
593 E T2c, T2f, T2i, T2j;
|
cannam@167
|
594 T2c = FMA(KP956940335, T2a, KP290284677 * T2b);
|
cannam@167
|
595 T2f = FNMS(KP290284677, T2e, KP956940335 * T2d);
|
cannam@167
|
596 T2g = T2c + T2f;
|
cannam@167
|
597 T2J = T2f - T2c;
|
cannam@167
|
598 T2i = FMA(KP290284677, T2d, KP956940335 * T2e);
|
cannam@167
|
599 T2j = FNMS(KP290284677, T2a, KP956940335 * T2b);
|
cannam@167
|
600 T2k = T2i - T2j;
|
cannam@167
|
601 T2N = T2j + T2i;
|
cannam@167
|
602 }
|
cannam@167
|
603 Cr[WS(csr, 14)] = T29 - T2g;
|
cannam@167
|
604 Ci[WS(csi, 14)] = T2N - T2O;
|
cannam@167
|
605 Cr[WS(csr, 1)] = T29 + T2g;
|
cannam@167
|
606 Ci[WS(csi, 1)] = T2N + T2O;
|
cannam@167
|
607 Cr[WS(csr, 9)] = T2h - T2k;
|
cannam@167
|
608 Ci[WS(csi, 9)] = T2J - T2M;
|
cannam@167
|
609 Cr[WS(csr, 6)] = T2h + T2k;
|
cannam@167
|
610 Ci[WS(csi, 6)] = T2J + T2M;
|
cannam@167
|
611 }
|
cannam@167
|
612 {
|
cannam@167
|
613 E T1n, T1v, T2y, T2A, T1u, T2v, T1y, T2z;
|
cannam@167
|
614 {
|
cannam@167
|
615 E T1j, T1m, T2w, T2x;
|
cannam@167
|
616 T1j = T5 - Tc;
|
cannam@167
|
617 T1m = T1k + T1l;
|
cannam@167
|
618 T1n = T1j + T1m;
|
cannam@167
|
619 T1v = T1j - T1m;
|
cannam@167
|
620 T2w = Tv - Tm;
|
cannam@167
|
621 T2x = T2q - T2n;
|
cannam@167
|
622 T2y = T2w - T2x;
|
cannam@167
|
623 T2A = T2w + T2x;
|
cannam@167
|
624 }
|
cannam@167
|
625 {
|
cannam@167
|
626 E T1q, T1t, T1w, T1x;
|
cannam@167
|
627 T1q = FMA(KP773010453, T1o, KP634393284 * T1p);
|
cannam@167
|
628 T1t = FNMS(KP634393284, T1s, KP773010453 * T1r);
|
cannam@167
|
629 T1u = T1q + T1t;
|
cannam@167
|
630 T2v = T1t - T1q;
|
cannam@167
|
631 T1w = FMA(KP634393284, T1r, KP773010453 * T1s);
|
cannam@167
|
632 T1x = FNMS(KP634393284, T1o, KP773010453 * T1p);
|
cannam@167
|
633 T1y = T1w - T1x;
|
cannam@167
|
634 T2z = T1x + T1w;
|
cannam@167
|
635 }
|
cannam@167
|
636 Cr[WS(csr, 12)] = T1n - T1u;
|
cannam@167
|
637 Ci[WS(csi, 12)] = T2z - T2A;
|
cannam@167
|
638 Cr[WS(csr, 3)] = T1n + T1u;
|
cannam@167
|
639 Ci[WS(csi, 3)] = T2z + T2A;
|
cannam@167
|
640 Cr[WS(csr, 11)] = T1v - T1y;
|
cannam@167
|
641 Ci[WS(csi, 11)] = T2v - T2y;
|
cannam@167
|
642 Cr[WS(csr, 4)] = T1v + T1y;
|
cannam@167
|
643 Ci[WS(csi, 4)] = T2v + T2y;
|
cannam@167
|
644 }
|
cannam@167
|
645 {
|
cannam@167
|
646 E T1L, T21, T2G, T2I, T20, T2H, T24, T2B;
|
cannam@167
|
647 {
|
cannam@167
|
648 E T1D, T1K, T2E, T2F;
|
cannam@167
|
649 T1D = T1z - T1C;
|
cannam@167
|
650 T1K = T1G - T1J;
|
cannam@167
|
651 T1L = T1D + T1K;
|
cannam@167
|
652 T21 = T1D - T1K;
|
cannam@167
|
653 T2E = T2C - T2D;
|
cannam@167
|
654 T2F = T26 + T27;
|
cannam@167
|
655 T2G = T2E - T2F;
|
cannam@167
|
656 T2I = T2F + T2E;
|
cannam@167
|
657 }
|
cannam@167
|
658 {
|
cannam@167
|
659 E T1S, T1Z, T22, T23;
|
cannam@167
|
660 T1S = FMA(KP881921264, T1O, KP471396736 * T1R);
|
cannam@167
|
661 T1Z = FMA(KP881921264, T1V, KP471396736 * T1Y);
|
cannam@167
|
662 T20 = T1S - T1Z;
|
cannam@167
|
663 T2H = T1S + T1Z;
|
cannam@167
|
664 T22 = FNMS(KP471396736, T1V, KP881921264 * T1Y);
|
cannam@167
|
665 T23 = FNMS(KP471396736, T1O, KP881921264 * T1R);
|
cannam@167
|
666 T24 = T22 - T23;
|
cannam@167
|
667 T2B = T23 + T22;
|
cannam@167
|
668 }
|
cannam@167
|
669 Cr[WS(csr, 13)] = T1L - T20;
|
cannam@167
|
670 Ci[WS(csi, 13)] = T2B - T2G;
|
cannam@167
|
671 Cr[WS(csr, 2)] = T1L + T20;
|
cannam@167
|
672 Ci[WS(csi, 2)] = T2B + T2G;
|
cannam@167
|
673 Cr[WS(csr, 10)] = T21 - T24;
|
cannam@167
|
674 Ci[WS(csi, 10)] = T2I - T2H;
|
cannam@167
|
675 Cr[WS(csr, 5)] = T21 + T24;
|
cannam@167
|
676 Ci[WS(csi, 5)] = -(T2H + T2I);
|
cannam@167
|
677 }
|
cannam@167
|
678 }
|
cannam@167
|
679 }
|
cannam@167
|
680 }
|
cannam@167
|
681
|
cannam@167
|
682 static const kr2c_desc desc = { 32, "r2cfII_32", {138, 46, 36, 0}, &GENUS };
|
cannam@167
|
683
|
cannam@167
|
684 void X(codelet_r2cfII_32) (planner *p) {
|
cannam@167
|
685 X(kr2c_register) (p, r2cfII_32, &desc);
|
cannam@167
|
686 }
|
cannam@167
|
687
|
cannam@167
|
688 #endif
|